How is the lifetime of SSDs that switch from SLC to MLC affected by the switch

Mark F wrote:
There are some SSDs that automatically switch from SLC to MLC when the
disk gets full.

Originally was found to be an undocumented feature in some drives;
if you allocated less than 1/2 or 1/3 of the drive to partitions, then
the firmware would run the drive effectively as SLC and things would
run faster with fewer very long delays and a longer lifetime.

Now there are some drives that are documented as being able to
dynamically switch parts of the drive from SLC to MLC (3 bits/cell in
the case of at least one drive type that I saw on the web.) (I didn't
read enough to determine if areas would be switched back to SLC
operation if space was freed up by TRIM.)

I could not tell would happen to the total data written
life of the drive it operated at SLC for a time and then switched to
MLC.

Does anyone know?

For example, 1 full disk write at SLC would be the same number
of cell writes as 3 full disk writes al MLC. How would this affect
the life of the drive when if had to permanently switch to MLC?

Spunds like you misunderstood something. Have a link?

Arno

On 10 Oct 2014 15:15:56 GMT, Arno wrote:

Mark F wrote:
There are some SSDs that automatically switch from SLC to MLC when the
disk gets full.

Originally was found to be an undocumented feature in some drives;
if you allocated less than 1/2 or 1/3 of the drive to partitions, then
the firmware would run the drive effectively as SLC and things would
run faster with fewer very long delays and a longer lifetime.

Now there are some drives that are documented as being able to
dynamically switch parts of the drive from SLC to MLC (3 bits/cell in
the case of at least one drive type that I saw on the web.) (I didn't
read enough to determine if areas would be switched back to SLC
operation if space was freed up by TRIM.)

I could not tell would happen to the total data written
life of the drive it operated at SLC for a time and then switched to
MLC.

Does anyone know?

For example, 1 full disk write at SLC would be the same number
of cell writes as 3 full disk writes al MLC. How would this affect
the life of the drive when if had to permanently switch to MLC?

Spunds like you misunderstood something. Have a link?
Yes, it turns out that I did misunderstand something. See
my added description below for the 5 cases that I have
identified.

Arno
There are many things around.
Case 1:
SLC caching (I don't have a link. I think from about 2013)
Typically documented as a feature.

Here there is dedicated SLC that is written to and
ultimately intended to me moved to the main arrays.
These cells are operated to run faster than the
main cells in their normal bit/cell mode.
(There could also be RAM cells in addition to or instead
of the faster running flash memory cells.)

Can also occur along with any of the other cases.


Case 2:
All data is stored in the main arrays as SLC and then
all is converted to more bits/cell if and when the drive determines
that more space is needed. Not a documented feature.

(I don't have a link. I think starting circa 2012)


Case 3:
Some main array cells are switched from SLC to MLC
as needed.

No long delay, may or may not be a documented feature

(I don't have a link. I think circa 2013.)

Case 4:
. A variable amount of space is operated as SLC and treated as
cache. The amount reserved depends on the storage in use.
The data is written directly to a more bits per cell a
area if the cache is full. The SLC cached stuff is always bit
in a more bits per cell area when the drive has a chance to
catch up. Samsung 840 EVO uses 3-bit/cell and calls it
"TurboWrite". This company white
paper is dated 2013 October 23:

https://www.samsung.com/it/business-...F_131023-0.pdf

I thought that the Micron M600 described in the writeup below
was what I am calling Case 5, but which I now think is Case 4:

http://www.micron.com/about/blogs/20...ance-and-value
and

http://www.micron.com/-/media/docume...tech_brief.pdf
The above is dated 2014 September 15

The page 5 of the Tech Brief briefly mentions the affect of
Dynamic Write Acceleration on the Write Amplification Factor.
However, the relative affect of writing as SLC compared to MLC,
which was my question, is not discussed.


The Micron M600 drives were reviewed in
http://www.anandtech.com/show/8528/m...da-placeholder
By Kristian Vatto (the "a" and "e" should have umlaut diacritic)
dated 2014 September 28
The review indicates that a Micron presentation indicated that the
data was cached in SLC but was always moved to MLC as time
permitted like my case 4.
NOTE: I originally thought that the Micron M600 operated as Case 5
when TRIM commands were issued before the caching mechanism
had done the SLC to MLC catchup.


Case 5:
A variable amount of space is operated as SLC and moved
to areas with more bits/cell as space is needed.
The difference from "TurboWrite" and the like is that the SLC is
not cache; the data is stored SLC until more space is needed.

I thought the Micron M600 drives were an example of Case 5,
but as per the "NOTE" in Case 4, I now think the M600's only
do caching. At this point I am not sure that I have seen
descriptions of any drives that actually fall in Case 5.


My question about lifetime referred to Case 5,
but now I see that it also be should be considered for Case 4.
and even Cases 2 and 3.

Mark F wrote:
On 10 Oct 2014 15:15:56 GMT, Arno wrote:

Mark F wrote:
There are some SSDs that automatically switch from SLC to MLC when the
disk gets full.

Originally was found to be an undocumented feature in some drives;
if you allocated less than 1/2 or 1/3 of the drive to partitions, then
the firmware would run the drive effectively as SLC and things would
run faster with fewer very long delays and a longer lifetime.

Now there are some drives that are documented as being able to
dynamically switch parts of the drive from SLC to MLC (3 bits/cell in
the case of at least one drive type that I saw on the web.) (I didn't
read enough to determine if areas would be switched back to SLC
operation if space was freed up by TRIM.)

I could not tell would happen to the total data written
life of the drive it operated at SLC for a time and then switched to
MLC.

Does anyone know?

For example, 1 full disk write at SLC would be the same number
of cell writes as 3 full disk writes al MLC. How would this affect
the life of the drive when if had to permanently switch to MLC?

Spunds like you misunderstood something. Have a link?
Yes, it turns out that I did misunderstand something. See
my added description below for the 5 cases that I have
identified.

Arno
There are many things around.
Case 1:
SLC caching (I don't have a link. I think from about 2013)
Typically documented as a feature.

Here there is dedicated SLC that is written to and
ultimately intended to me moved to the main arrays.
These cells are operated to run faster than the
main cells in their normal bit/cell mode.
(There could also be RAM cells in addition to or instead
of the faster running flash memory cells.)

Can also occur along with any of the other cases.


Case 2:
All data is stored in the main arrays as SLC and then
all is converted to more bits/cell if and when the drive determines
that more space is needed. Not a documented feature.

(I don't have a link. I think starting circa 2012)


Case 3:
Some main array cells are switched from SLC to MLC
as needed.

No long delay, may or may not be a documented feature

(I don't have a link. I think circa 2013.)

Case 4:
. A variable amount of space is operated as SLC and treated as
cache. The amount reserved depends on the storage in use.
The data is written directly to a more bits per cell a
area if the cache is full. The SLC cached stuff is always bit
in a more bits per cell area when the drive has a chance to
catch up. Samsung 840 EVO uses 3-bit/cell and calls it
"TurboWrite". This company white
paper is dated 2013 October 23:

https://www.samsung.com/it/business-...F_131023-0.pdf

I thought that the Micron M600 described in the writeup below
was what I am calling Case 5, but which I now think is Case 4:

http://www.micron.com/about/blogs/20...ance-and-value
and

http://www.micron.com/-/media/docume...tech_brief.pdf
The above is dated 2014 September 15

The page 5 of the Tech Brief briefly mentions the affect of
Dynamic Write Acceleration on the Write Amplification Factor.
However, the relative affect of writing as SLC compared to MLC,
which was my question, is not discussed.


The Micron M600 drives were reviewed in
http://www.anandtech.com/show/8528/m...da-placeholder
By Kristian Vatto (the "a" and "e" should have umlaut diacritic)
dated 2014 September 28
The review indicates that a Micron presentation indicated that the
data was cached in SLC but was always moved to MLC as time
permitted like my case 4.
NOTE: I originally thought that the Micron M600 operated as Case 5
when TRIM commands were issued before the caching mechanism
had done the SLC to MLC catchup.


Case 5:
A variable amount of space is operated as SLC and moved
to areas with more bits/cell as space is needed.
The difference from "TurboWrite" and the like is that the SLC is
not cache; the data is stored SLC until more space is needed.

I thought the Micron M600 drives were an example of Case 5,
but as per the "NOTE" in Case 4, I now think the M600's only
do caching. At this point I am not sure that I have seen
descriptions of any drives that actually fall in Case 5.


My question about lifetime referred to Case 5,
but now I see that it also be should be considered for Case 4.
and even Cases 2 and 3.

Ok, I see. First, the damage do a FLASH cell is mainly not
caused by writing, it is caused by erasing. Of course in order
to write a cell multiple times, you have to erase it multiple
times.

But then it gets complicated:
SLC cells only have two levels: Something in there and nothing
in there in therms of charge. That is very robust. For example,
you can charge a cell to some value "a" on writing a cell, and
you can reliably detect this even when the charge has droppend
to 10% or less. This also means that SLC cells can sustain quite
some damage and still be reliable.

So for SLC cells you do not have to be very careful how much
carge to put in on writing. You still want to limit it, as
otherwise you do get much more damage from erasing the cells.


MLC, on the other hand has 4 levels and TLC has 8 levels.
This has two effects: First, the levels put in there need
to be verified as smaller changes can change the value.
Second, smaller drift due to cell damage is tolerable, and
hence less cell damage is tolerabe.

So there are two factors:
- SLC can sustain more erasures and still work reliably. This
can be a factor of 10 for MLC and a factor for 100 TLC
and even more.
- MLC only writes half the number of cells, but writes 3 out of
4 times, TLS only writes 1/3 the number of cells, but writes
7 out of 8 times, while SLC only writes a cell for 50% of the
time. Why is this relevant? Simple: Erasing does not damage FLASH
cells, erasing charged cells does, and somewhat proportional to
the charge in there.
As on average a cell gets charged 50%, the damage roughly is
the same per erasure per cell. So MLC suffers half and TLC
suffers 1/3. But MLC can only tolerate 10% the writes and
TLS can only tolerate 1% the writes.

Bottom line is that using Cells for SLC makes them untuitable
to be used as MLS or TLC later on, as very few SLC erasures
will drive the cells over the damage limit for MLC or TLC.

Of course, that is not the whole truth: If you run SLCs witout
the additional reliability larger charges give you, you do less
damage to them. They would still be fast, because if you only
have two levels, you do not need the verify thet modern MLC and
TLC needs (at least after a number of cycles).

An just to make matters even worse: Modern SSDs do periodic
refreshes of cells that have gotten weaker, but only if they
are powered. If they sit on a shelf the data is not getting
refreshed.

The bottom line is this: There is no way to tell
the impact of this without knowning the details.
You can, to a degree, fall back on warranty periods
and the conditions they are specified under.

So what do you do? Simple:
Make sure you do periodic full scans of the data. This is also
somethign you needed to do for magnetic disks in order to have
reliability. A long SMART selftest (with alerting in case
something fails!) every 14 days or so is highly advisable.
And make sure that when the SSD fails to have backup
or RAID and be able to do a low-cost recovery.

Arno